Difference between revisions of "Toward hyperdimensional active perception: learning compressed sensorimotor control by demonstration"

Revision as of 16:05, 29 October 2019

Description

Hyperdimensional (HD) computing [1] is a brain-inspired computing paradigm based on representing information with hypervectors thousands of dimensions). Hypervectors are holographic and (pseudo)random with independent and identically distributed (i.i.d.) components. Principles of HD computing allow to implement efficient learning models as well as universal computing, e.g., emulating finite state automata [2]. In this project, your goal would be to improve algorithms, and implementations based on HD computing for robotics applications, especially with the dataset studies in [3].

There are three main objectives for this project.

• Implementing the architecture of [3] on an embedded platform such as Jetson TX2, and compare the performance of both CNN and HD methods. The HD method should be improved toward a more efficient realization.
• Implementing the DVS sensory representation in RTL, and assess how much compression can be done for an acceptable reconstruction of events.
• Evaluating the methods with other datasets [4].

Status: In progress

Edoardo Mello Rella

Supervision: Michael Hersche, Alfio Di Mauro Abbas Rahimi

Prerequisites

• Machine Learning
• Python Programming
• Verilog Programming

Character

20% Theory

80% Programming

Professor

Luca Benini

↑ top

Literature

• [1] Kanerva, Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors
• [2] Osipov et al., Associative synthesis of finite state automata model of a controlled object with hyperdimensional computing
• [3] Mitrokhin et al., Learning sensorimotor control with neuromorphic sensors: Toward hyperdimensional active perception
• [4] Maqueda et al., Event-Based Vision Meets Deep Learning on Steering Prediction for Self-Driving Cars